---
title: "More Data, Fewer Diacritics — Musleh 2026"
description: "Arabic speech synthesis has two problems. There is little public training audio, and written Arabic usually omits diacritics, the short-vowel marks that…"
authors: ["Ahmed Musleh", "Yifan Zhang", "Kareem Darwish"]
year: 2026
venue: "arXiv:2603.01622"
source: "https://arxiv.org/abs/2603.01622"
doi: "10.48550/arxiv.2603.01622"
arxiv: "2603.01622"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/musleh-2026-more-data-fewer-diacritics-scaling/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/musleh-2026-more-data-fewer-diacritics-scaling/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/7/) / More Data, Fewer Diacritics: Scaling Arabic TTS

# More Data, Fewer Diacritics: Scaling Arabic TTS

**Authors:** Ahmed Musleh, [Yifan Zhang](https://md.research.elamur.ai/free/en/authors/yifan-zhang/), [Kareem Darwish](https://md.research.elamur.ai/free/en/authors/kareem-darwish/) · **Year:** 2026 · **Venue:** arXiv:2603.01622 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2603.01622) · [PDF](https://arxiv.org/pdf/2603.01622) · [DOI 10.48550/arxiv.2603.01622](https://doi.org/10.48550/arxiv.2603.01622) · [arXiv 2603.01622](https://arxiv.org/abs/2603.01622) · [OpenAlex](https://openalex.org/W7133315578)

**Citations:** 1 citations (openalex-snapshot · 2026-10-14) · 1 in our library, of the 140884 papers whose bibliographies we read

## Summary
Arabic speech synthesis has two problems. There is little public training audio, and written Arabic usually omits diacritics, the short-vowel marks that determine pronunciation. Existing public corpora are small and fully diacritized; the largest has 83 hours, most of it synthetic. The authors gathered about 20k hours of long-form audio shared on a social media platform. They processed it with a fully automatic chain: Silero VAD for segmentation, Fanar ASR for transcripts, a re-implemented character-level diacritizer, and a noise-and-length filter. This left about 4,000 hours of 3–15-second clips, each with both plain and diacritized transcripts.

They trained eight F5-TTS voice-cloning models on 100, 1,000 and 4,000 hours, with and without diacritics. Diacritized input won at every size. However, more data substantially offset the missing marks: the undiacritized 1,000-hour model beat both diacritized 100-hour models. The practical message is that collecting more data may matter more than perfect diacritization. The authors plan to release a model that reads raw Arabic text.

## Key points
- **Pipeline:**
  - Silero VAD segments the audio.
  - Fanar ASR (conformer, roughly 15K training hours of Arabic and English) produces transcripts.
  - A re-implementation of the Elmallah et al. (2024) diacritizer adds diacritics. It was trained on a 4M-word in-house MSA corpus and scores 5.5% WER on WikiNews.
  - Filtering drops a segment if the noise in the 1 s before or after it exceeds -30 dB, or if the clip is outside 3–15 s.
- **Training sets:** eight sets in total.
  - 4,000 h and 1,000 h, each with and without diacritics.
  - Two 100 h subsets, one with maximum and one with minimum speaker diversity, each with and without diacritics.
  - Each model was trained for 600k steps (DiT with 22 layers, dim 1024, Vocos vocoder).
- **Test set:** 59 MSA speakers found with Pyannote diarization, 11 clips each (1 cloning prompt and 10 targets). The source files of these speakers were removed from training.
- **Results (WER / SpeechBERTScore):**

  | Size | Diacritics | Diversity | WER | SpeechBERTScore |
  |---|---|---|---|---|
  | 4,000 h | yes | – | 1.67 | 0.739 |
  | 1,000 h | yes | – | 2.55 | 0.729 |
  | 1,000 h | no | – | 3.12 | 0.726 |
  | 4,000 h | no | – | 3.18 | 0.732 |
  | 100 h | yes | max | 3.48 | 0.719 |
  | 100 h | yes | min | 3.62 | 0.712 |
  | 100 h | no | max | 4.06 | 0.704 |
  | 100 h | no | min | 5.20 | 0.706 |

  - The two metrics rank the undiacritized 4,000 h model differently relative to the 1,000 h models.
  - Note: in raw WER, the gap between diacritized and plain input is actually largest at 4,000 h.
- **Limitations the authors acknowledge:**
  - No human evaluation.
  - The same ASR family both built the data and scored WER.
  - Only MSA was tested; no dialect analysis.

## Interesting ideas
- Several imperfect automatic annotators, chained with cheap physical filters (surrounding noise, clip length), can turn raw unlabeled material into a large silver corpus whose errors matter less as it grows. — For Mova (learn): Our witness-agreement silver pipeline could likewise add cheap, model-independent sentence filters (length, encoding noise, punctuation oddities). It could then measure, as here, whether silver volume outruns silver noise.
- The paper runs a factorial ablation: an explicit intermediate annotation (diacritics) is turned on or off at several data scales, to see when a model absorbs that layer implicitly. — For Mova (tokens): We could train the tagger with and without explicit morphological features at several silver sizes. This would show where hand-built morphology still pays off and where context alone suffices.
- A rough estimate converts the diacritizer's training word count into equivalent audio hours. This predicts the data size at which the end model could learn the hidden layer by itself. — For Mova (learn): Before building a dedicated component, we could compare its training-data size with the word volume flowing through a downstream model. This would tell us whether the downstream model might subsume it.
- To avoid contamination, entire source files that contain test speakers are excluded from training, not just the test clips. The authors also flag the bias of scoring with the same ASR that produced the labels. — For Mova (learn): Treebank and silver splits should hold out by document. Silver-trained parsers should be scored by a witness that did not help create the silver data, which serves as a built-in negative control.
- Two training subsets of equal size, one with maximal and one with minimal speaker diversity, isolate the effect of diversity from the effect of volume. — For Mova (tree): Building equal-size gparse training subsets with maximal and minimal genre or source diversity would show whether parser gains come from variety or from sheer token count.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{musleh2026more,
  title = {{More Data, Fewer Diacritics: Scaling Arabic TTS}},
  author = {Musleh, Ahmed and Zhang, Yifan and Darwish, Kareem},
  year = {2026},
  doi = {10.48550/arxiv.2603.01622},
  eprint = {2603.01622},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2603.01622}
}
```

- [← previous Modular Neural Machine Translation with a Semantic Pivot - Pilot Study Using AMR](https://md.research.elamur.ai/free/en/papers/gao-2026-modular-neural-machine-translation-semantic/)
- [next → More on the Arabic Object Marker iyyā: Implications for the Origin of the Semitic Notae Accusativi](https://md.research.elamur.ai/free/en/papers/wilmsen-2026-more-arabic-object-marker-iyya/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/7/)

## References

22 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Ahmed Abdelali, Nadir Durrani, Cenk Demiroglu, Fahim Dalvi, Hamdy Mubarak, and Kareem Darwish. 2022. Natiq: An end-to-end text-to-speech system for arabic. arXiv preprint arXiv:2206.07373. [↗](https://arxiv.org/abs/2206.07373)
2. Ibrahim Almosallam, Atheer AlKhalifa, Mansour Al-Ghamdi, Mohamed I Alkanhal, and Ashraf Alkhairy. 2013. Sassc: a standard arabic single speaker corpus. In SSW, pages 249–253. [find](https://scholar.google.com/scholar?q=Ibrahim+Almosallam%2C+Atheer+AlKhalifa%2C+Mansour+Al%2DGhamdi%2C+Mohamed+I+Alkanhal%2C+and+Ashraf+Alkhairy%2E+2013%2E+Sassc%3A+a+standard+arabic+single+speaker+corpus%2E+In+SSW%2C+pages+249%E2%80%93253%2E)
3. Hervé Bredin. 2023. pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe. In Proc. INTERSPEECH 2023. [find](https://scholar.google.com/scholar?q=Herv%C3%A9+Bredin%2E+2023%2E+pyannote%2Eaudio+2%2E1+speaker+diarization+pipeline%3A+principle%2C+benchmark%2C+and+recipe%2E+In+Proc%2E+INTERSPEECH+2023%2E)
4. Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao, Kai Yu, and Xie Chen. 2024. F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching. arXiv preprint arXiv:2410.06885. [↗](https://arxiv.org/abs/2410.06885)
5. Kareem Darwish, Hamdy Mubarak, and Ahmed Abdelali. 2017. Arabic diacritization: Stats, rules, and hacks. In Proceedings of the third Arabic natural language processing workshop, pages 9–17. [find](https://scholar.google.com/scholar?q=Kareem+Darwish%2C+Hamdy+Mubarak%2C+and+Ahmed+Abdelali%2E+2017%2E+Arabic+diacritization%3A+Stats%2C+rules%2C+and+hacks%2E+In+Proceedings+of+the+third+Arabic+natural+language+processing+workshop%2C+pages+9%E2%80%9317%2E) [our card](https://md.research.elamur.ai/free/en/papers/aclbulk-2017-darwish-arabic-diacritization-stats-rules-hacks/)
6. Muhammad Morsy Elmallah, Mahmoud Reda, Kareem Darwish, Abdelrahman El-Sheikh, Ashraf Hatim Elneima, Murtadha Aljubran, Nouf Alsaeed, Reem Mohammed, and Mohamed Al-Badrashiny. 2024. Arabic diacritization using morphologically informed character-level model. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 1446–1454. [find](https://scholar.google.com/scholar?q=Muhammad+Morsy+Elmallah%2C+Mahmoud+Reda%2C+Kareem+Darwish%2C+Abdelrahman+El%2DSheikh%2C+Ashraf+Hatim+Elneima%2C+Murtadha+Aljubran%2C+Nouf+Alsaeed%2C+Reem+Mohammed%2C+and+Mohamed+Al%2DBadrashiny%2E+2024%2E+Arabic+diacritizati) [our card](https://md.research.elamur.ai/free/en/papers/coling-2024-elmallah-arabic-diacritization-morphologically-informed-character/)
7. Sefik Emre Eskimez, Xiaofei Wang, Manthan Thakker, Canrun Li, Chung-Hsien Tsai, Zhen Xiao, Hemin Yang, Zirun Zhu, Min Tang, Xu Tan, and 1 others. 2024. E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts. In 2024 IEEE Spoken Language Technology Workshop (SLT), pages 682– 689. IEEE. [find](https://scholar.google.com/scholar?q=Sefik+Emre+Eskimez%2C+Xiaofei+Wang%2C+Manthan+Thakker%2C+Canrun+Li%2C+Chung%2DHsien+Tsai%2C+Zhen+Xiao%2C+Hemin+Yang%2C+Zirun+Zhu%2C+Min+Tang%2C+Xu+Tan%2C+and+1+others%2E+2024%2E+E2+tts%3A+Embarrassingly+easy+fully+non%2Dautoregres)
8. Nawar Halabi and 1 others. 2016. Arabic speech corpus. Oxford Text Archive Core Collection. [find](https://scholar.google.com/scholar?q=Nawar+Halabi+and+1+others%2E+2016%2E+Arabic+speech+corpus%2E+Oxford+Text+Archive+Core+Collection%2E)
9. Shijia Liao, Yuxuan Wang, Tianyu Li, Yifan Cheng, Ruoyi Zhang, Rongzhi Zhou, and Yijin Xing. 2024. Fish-speech: Leveraging large language models for advanced multilingual text-to-speech synthesis. Preprint, arXiv:2411.01156. mrfakename, Vaibhav Srivastav, Clémentine Fourrier, Lucain Pouget, Yoach Lacombe, main, and Sanchit Gandhi. 2024. Text to speech arena. https://huggingface.co/spaces/TTS-AGI/TTS-Arena. [↗](https://arxiv.org/abs/2411.01156)
10. Hawau Olamide Toyin, Rufael Marew, Humaid Alblooshi, Samar M Magdy, and Hanan Aldarmaki. 2025. Arvoice: A multi-speaker dataset for arabic speech synthesis. arXiv e-prints, pages arXiv–2505. [find](https://scholar.google.com/scholar?q=Hawau+Olamide+Toyin%2C+Rufael+Marew%2C+Humaid+Alblooshi%2C+Samar+M+Magdy%2C+and+Hanan+Aldarmaki%2E+2025%2E+Arvoice%3A+A+multi%2Dspeaker+dataset+for+arabic+speech+synthesis%2E+arXiv+e%2Dprints%2C+pages+arXiv%E2%80%932505%2E) [our card](https://md.research.elamur.ai/free/en/papers/toyin-2025-arvoice-multi-speaker-dataset-arabic/)
11. Alexis Plaquet and Hervé Bredin. 2023. Powerset multi-class cross entropy loss for neural speaker diarization. In Proc. INTERSPEECH 2023. [find](https://scholar.google.com/scholar?q=Alexis+Plaquet+and+Herv%C3%A9+Bredin%2E+2023%2E+Powerset+multi%2Dclass+cross+entropy+loss+for+neural+speaker+diarization%2E+In+Proc%2E+INTERSPEECH+2023%2E)
12. Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. 2021. Grad-tts: A diffusion probabilistic model for text-to-speech. In International conference on machine learning, pages 8599–8608. PMLR. [find](https://scholar.google.com/scholar?q=Vadim+Popov%2C+Ivan+Vovk%2C+Vladimir+Gogoryan%2C+Tasnima+Sadekova%2C+and+Mikhail+Kudinov%2E+2021%2E+Grad%2Dtts%3A+A+diffusion+probabilistic+model+for+text%2Dto%2Dspeech%2E+In+International+conference+on+machine+learning%2C+p)
13. Takaaki Saeki, Soumi Maiti, Shinnosuke Takamichi, Shinji Watanabe, and Hiroshi Saruwatari. 2024. SpeechBERTScore: Reference-aware automatic evaluation of speech generation leveraging nlp evaluation metrics. In Interspeech 2024, pages 4943–4947. [find](https://scholar.google.com/scholar?q=Takaaki+Saeki%2C+Soumi+Maiti%2C+Shinnosuke+Takamichi%2C+Shinji+Watanabe%2C+and+Hiroshi+Saruwatari%2E+2024%2E+SpeechBERTScore%3A+Reference%2Daware+automatic+evaluation+of+speech+generation+leveraging+nlp+evaluation+me)
14. Hubert Siuzdak. 2023. Vocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis. arXiv preprint arXiv:2306.00814. [↗](https://arxiv.org/abs/2306.00814)
15. Fanar Team, Ummar Abbas, Mohammad Shahmeer Ahmad, Firoj Alam, Enes Altinisik, Ehsannedin Asgari, Yazan Boshmaf, Sabri Boughorbel, Sanjay Chawla, Shammur Chowdhury, and 1 others. 2025. Fanar: An arabic-centric multimodal generative ai platform. arXiv preprint arXiv:2501.13944. [↗](https://arxiv.org/abs/2501.13944) [our card](https://md.research.elamur.ai/free/en/papers/abbas-2025-fanar-arabic-centric-multimodal-generative/)
16. Silero Team. 2024. Silero vad: pre-trained enterprise-grade voice activity detector (vad), number detector and language classifier. [https://github.com/](https://github.com/) snakers4/silero-vad. [↗](https://github.com/)
17. Andrew J Hunt and Alan W Black. 1996. Unit selection in a concatenative speech synthesis system using a large speech database. In 1996 IEEE international conference on acoustics, speech, and signal processing conference proceedings, volume 1, pages 373–376. IEEE. [find](https://scholar.google.com/scholar?q=Andrew+J+Hunt+and+Alan+W+Black%2E+1996%2E+Unit+selection+in+a+concatenative+speech+synthesis+system+using+a+large+speech+database%2E+In+1996+IEEE+international+conference+on+acoustics%2C+speech%2C+and+signal+pr)
18. Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, and 1 others. 2017. Tacotron: Towards end-to-end speech synthesis. arXiv preprint arXiv:1703.10135. [↗](https://arxiv.org/abs/1703.10135)
19. Ajinkya Kulkarni, Atharva Kulkarni, Sara Abedal-monem Mohammad Shatnawi, and Hanan Aldarmaki. 2023. Clartts: An open-source classical arabic text-to-speech corpus. arXiv preprint arXiv:2303.00069. [↗](https://arxiv.org/abs/2303.00069) [our card](https://md.research.elamur.ai/free/en/papers/kulkarni-2023-clartts-open-source-classical-arabic/)
20. Jiahong Yuan, Neville Ryant, Mark Liberman, Andreas Stolcke, Vikramjit Mitra, and Wen Wang. 2013. Automatic phonetic segmentation using boundary models. In Interspeech, pages 2306–2310. [find](https://scholar.google.com/scholar?q=Jiahong+Yuan%2C+Neville+Ryant%2C+Mark+Liberman%2C+Andreas+Stolcke%2C+Vikramjit+Mitra%2C+and+Wen+Wang%2E+2013%2E+Automatic+phonetic+segmentation+using+boundary+models%2E+In+Interspeech%2C+pages+2306%E2%80%932310%2E)
21. Yinghao Aaron Li, Cong Han, Vinay Raghavan, Gavin Mischler, and Nima Mesgarani. 2023. Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models. Advances in Neural Information Processing Systems, 36:19594–19621. [find](https://scholar.google.com/scholar?q=Yinghao+Aaron+Li%2C+Cong+Han%2C+Vinay+Raghavan%2C+Gavin+Mischler%2C+and+Nima+Mesgarani%2E+2023%2E+Styletts+2%3A+Towards+human%2Dlevel+text%2Dto%2Dspeech+through+style+diffusion+and+adversarial+training+with+large+speech+)
22. Heiga Zen, Takashi Nose, Junichi Yamagishi, Shinji Sako, Takashi Masuko, Alan W Black, and Keiichi Tokuda. 2007. The hmm-based speech synthesis system (hts) version 2.0. SSW, 6:294–299. [find](https://scholar.google.com/scholar?q=Heiga+Zen%2C+Takashi+Nose%2C+Junichi+Yamagishi%2C+Shinji+Sako%2C+Takashi+Masuko%2C+Alan+W+Black%2C+and+Keiichi+Tokuda%2E+2007%2E+The+hmm%2Dbased+speech+synthesis+system+%28hts%29+version+2%2E0%2E+SSW%2C+6%3A294%E2%80%93299%2E)